Semantic Segmentation of Indoor Point Clouds with the Channel Attention Mechanism

Abstract

In recent years, the semantic segmentation of 3D point cloud has received increasing attention the field of computer vision, because 3D point cloud can better reflect our 3D space. Because of the unstructured and disordered characteristics of 3D point cloud data, semantic segmentation of point cloud is still a difficult task. Our network automatically learns the importance of feature channels by adding a channel attention module, which enables the network to obtain better training results. After the channel attention module is fused, the important channels in the features are enhanced, and the unimportant channels are suppressed, making the network training more efficient. In this paper, we propose an indoor point cloud semantic segmentation method combined with channel attention mechanism. The experimental results show that our method achieves better results than other methods.